Proprietary AI Tools

The AI starts working after the merge: Google Labs' Stitch Loop hunts problems, writes the fix, and hands you a pull request

November 20, 20260 views阅读中文原文
The AI starts working after the merge: Google Labs' Stitch Loop hunts problems, writes the fix, and hands you a pull request

TL;DR

On November 18, Google Labs introduced Stitch Loop, an experiment it describes as a long-running product agent. It connects to your GitHub repository and your live production metrics, works out what is dragging the product down, writes the fix inside an isolated sandbox, and opens a GitHub pull request for you to review.

The distinction from today's coding agents is simple. They clock out the moment you hit merge; Stitch Loop clocks in at that moment. Google's own wording on X: "Today's coding agents stop when code ships."

It is not a bot that pushes to production on its own. The company states plainly that every change waits on a human reviewing the pull request. Nothing reaches production until you approve it.

It is also a closed, commercial product in experimental form: the models are not open weights, and there is no self-hosted option. For now the only way in is the waitlist, open worldwide, with Google AI Ultra subscribers getting priority. Pricing and general availability have not been announced.

What it actually does

The pipeline is what determines whether this fits your team, so start there.

Setup means connecting context: the GitHub repository, plus product signals such as logs, user behavior, and user feedback. Google offers three routes — more than 180 integrations, custom MCP servers, or local CLI uploads. The agent is not meant to guess; every claim it makes has to be grounded in that data.

After setup it runs loops: a repeating process aimed at a measurable goal you define. Each loop moves through four stages.

  1. Suggestions. It surfaces opportunities to move the goal forward, backed by evidence from the product rather than a generic to-do list.
  2. Solutions. Each suggestion arrives with one or more approaches, including trade-offs and scoping, plus visual design options where the change is interface-facing.
  3. Build. Once you pick a direction, it writes the code in an isolated cloud sandbox and opens a GitHub pull request. When relevant, the PR carries screenshots or video of what it changed.
  4. Reflect. It keeps watching live data, learning from each positive and negative result and feeding that into the next round. Reaching the goal does not end the job — it stays on watch in case the metric slips.

The interesting part is not that an AI can write code, which stopped being novel a while ago. It is that the discovery, the proposal, the implementation, and the check on whether it worked are one closed loop, with evidence recorded at every stage.

Three loops in the box, and the ones you add

Three loops ship by default, aimed at the parts small teams most often let rot: cybersecurity, latency, and accessibility.

The security loop deserves detail. It combines the cyber capabilities of Gemini 4 Carbon with CodeMender, Google DeepMind's autonomous agent for improving code security. Together they scan the codebase for vulnerabilities before those become incidents, flag each finding with evidence, and open a PR with a proposed fix.

Latency and accessibility work the same way: one watches response speed, the other checks against WCAG accessibility standards. You can also build your own loops; Google's example is lifting checkout conversion — exactly the kind of goal tied directly to revenue where a looping agent makes sense.

How to use it: join the queue, don't plan a deployment

This needs saying clearly so nobody assumes it can go into production today.

  1. Join the waitlist first. It is at stitch.google.com/loop, open globally. A Google AI Ultra subscription is the only accelerator, and it grants priority rather than immediate access.
  2. Decide which number you want moved. The whole design revolves around measurable outcomes. If your team cannot say which dropped metric counts as a problem, the agent has nothing useful to suggest and you have no basis for approving its PRs.
  3. Fix your data plumbing. GitHub permissions, logs, behavioral data. The 180-plus integrations and MCP support exist for this step. If the data cannot get in, you are left with an ordinary assistant that can only read code.
  4. Treat it as a service layer, not a replacement. Coding agents such as Google Antigravity can pull persistent product context from Stitch Loop through the Stitch CLI and MCP. In Slack, an @stitch bot delivers daily insight digests and can create new loops. In other words, it can own the finding-and-judging half while your existing toolchain does the rest.

Where it is not a good fit

This is a Labs experiment, not a generally available commercial product, and that premise drives everything below.

  • No pricing and no GA date. Japanese outlet TechnoEdge noted explicitly in its coverage that pricing and general availability are unknown. Teams that need to budget a rollout cannot do the math yet.
  • Not for teams that will not review PRs. Its output is always a pull request, never a direct commit. If what you want is unattended automatic repair, the product design is a flat refusal.
  • Not for projects without a test baseline or stable metrics. Every step demands evidence and a measurable goal. Without a baseline it can only guess.
  • Strong compliance environments need extra review. Handing your repository, logs, and user behavior data to a cloud agent is a security review on its own, regardless of who ships it.
  • It will not replace engineers. It handles the post-launch chores that never make it into a sprint plan. It does not make product decisions.

Pitfalls: four easy misreadings

  • "It will change production by itself." No. Sandbox, pull request, human approval. The blog stresses three separate times that you make the final call.
  • "More PRs is better." The opposite. An agent running around the clock with a vague goal produces a PR flood, the team cannot keep up, and everything gets ignored. Start with one or two loops and add more once they prove useful.
  • "Connect it and conversion goes up." It optimizes the metric you define. Define the wrong one and it will steer the product off course very efficiently.
  • "It ships screenshots, so I don't need to read the code." The visuals help you judge direction fast. They do not transfer the responsibility of the person clicking merge.

Cost and time

  • Money. Product pricing is unpublished; priority access is tied to a Google AI Ultra subscription, which is an existing cost. Inference runs on Google's side, so there is no GPU bill of your own.
  • People. The real cost is review. Every PR needs someone to read it, judge it, and own the merge. That is a core design constraint, not an optional step.
  • Time. Wiring up GitHub and data sources is an hours-long job; the waitlist is an unknown. If you want this to become a process, settle internally which loops you want, which metric they target, and who approves PRs — then you can start the day access lands.

Who it suits: from "can we use it" to "should we"

  • Good fit. Teams of three to five maintaining a live product with real users; teams carrying a backlog of security findings, slow pages, and accessibility gaps that never get scheduled; teams that already have a metrics dashboard and a PR review habit.
  • Worth watching. Larger organizations can trial it on a peripheral product to judge output quality before rolling it out.
  • Not yet. Products that have not launched and have no user data, since it optimizes the after-launch phase; organizations with hard constraints on code hosting and data residency; anyone expecting one sentence to fix a live incident.

Next steps

Step one, available today: join the waitlist at stitch.google.com/loop, and read the four loop examples in the launch post to see whether any map onto a real pain point.

Step two, align on one metric internally. Write down the sentence "if this number improves 10 percent, we win." If nobody can write it, do not apply yet.

Step three, name the person who owns PR review. The most common way a Stitch Loop rollout fails is not model quality — it is that nobody wants to review a bot's pull requests every day.

The pitch is modest when you strip the marketing: take the after-launch work nobody wants to do but somebody has to, hand it to an agent that never logs off, and keep the human finger on the merge button.

This article is compiled from multiple media reports, alongside Google's official blog and X post; timing and feature details follow the official account.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...